Under the potential outcomes framework, causal effects are defined as comparisons between potential outcomes under treatment and control. To infer causal effects from randomized experiments, Neyman proposed to test the null hypothesis of zero average causal effect (Neyman’s null), and Fisher proposed to test the null hypothesis of zero individual causal effect (Fisher’s null). Although the subtle difference between Neyman’s null and Fisher’s null has caused a lot of controversies and confusions for both theoretical and practical statisticians, a careful comparison between the two approaches has been lacking in the literature for more than eighty years. We fill this historical gap by making a theoretical comparison between them and highlighting an intriguing paradox that has not been recognized by previous researchers. Logically, Fisher’s null implies Neyman’s null. It is therefore surprising that, in actual completely randomized experiments, rejection of Neyman’s null does not imply rejection of Fisher’s null for many realistic situations, including the case with constant causal effect. Furthermore, we show that this paradox also exists in other commonly-used experiments, such as stratified experiments, matched-pair experiments and factorial experiments. Asymptotic analyses, numerical examples and real data examples all support this surprising phenomenon. Besides its historical and theoretical importance, this paradox also leads to useful practical implications for modern researchers.

Although both Fisher’s and Neyman’s tests are for testing “no treatment effects,” they both test fundamentally different null hypotheses. While Neyman’s null concerns the average casual effect, Fisher’s null focuses on the individual causal effect. When conducting a test, researchers need to understand what is really being tested and what underlying assumptions are being made. If these fundamental issues are not fully appreciated, dubious conclusions regarding causal effects can be made.

This is a contribution to the discussion of the interesting paper by Ding [Statist. Sci. 32 (2017) 331–345], which contrasts approaches attributed to Neyman and Fisher. I believe that Fisher’s usual assumption was unit-treatment additivity, rather than the “sharp null hypothesis” attributed to him. Fisher also developed the notion of interaction in factorial experiments. His explanation leads directly to the concept of marginality, which is essential for the interpretation of data from any factorial experiment.

A high quality logistic regression model contains various desirable properties: predictive power, interpretability, significance, robustness to error in data and sparsity, among others. To achieve these competing goals, modelers incorporate these properties iteratively as they hone in on a final model. In the period 1991–2015, algorithmic advances in Mixed-Integer Linear Optimization (MILO) coupled with hardware improvements have resulted in an astonishing 450 billion factor speedup in solving MILO problems. Motivated by this speedup, we propose modeling logistic regression problems algorithmically with a mixed integer nonlinear optimization (MINLO) approach in order to explicitly incorporate these properties in a joint, rather than sequential, fashion. The resulting MINLO is flexible and can be adjusted based on the needs of the modeler. Using both real and synthetic data, we demonstrate that the overall approach is generally applicable and provides high quality solutions in realistic timelines as well as a guarantee of suboptimality. When the MINLO is infeasible, we obtain a guarantee that imposing distinct statistical properties is simply not feasible.

Big Datasets are endemic, but are often notoriously difficult to analyse because of their size, heterogeneity and quality. The purpose of this paper is to open a discourse on the potential for modern decision theoretic optimal experimental design methods, which by their very nature have traditionally been applied prospectively, to improve the analysis of Big Data through retrospective designed sampling in order to answer particular questions of interest. By appealing to a range of examples, it is suggested that this perspective on Big Data modelling and analysis has the potential for wide generality and advantageous inferential and computational properties. We highlight current hurdles and open research questions surrounding efficient computational optimisation in using retrospective designs, and in part this paper is a call to the optimisation and experimental design communities to work together in the field of Big Data analysis.

The basic idea of importance sampling is to use independent samples from a proposal measure in order to approximate expectations with respect to a target measure. It is key to understand how many samples are required in order to guarantee accurate approximations. Intuitively, some notion of distance between the target and the proposal should determine the computational cost of the method. A major challenge is to quantify this distance in terms of parameters or statistics that are pertinent for the practitioner. The subject has attracted substantial interest from within a variety of communities. The objective of this paper is to overview and unify the resulting literature by creating an overarching framework. A general theory is presented, with a focus on the use of importance sampling in Bayesian inverse problems and filtering.

The propensity score is a common tool for estimating the causal effect of a binary treatment in observational data. In this setting, matching, subclassification, imputation or inverse probability weighting on the propensity score can reduce the initial covariate bias between the treatment and control groups. With more than two treatment options, however, estimation of causal effects requires additional assumptions and techniques, the implementations of which have varied across disciplines. This paper reviews current methods, and it identifies and contrasts the treatment effects that each one estimates. Additionally, we propose possible matching techniques for use with multiple, nominal categorical treatments, and use simulations to show how such algorithms can yield improved covariate similarity between those in the matched sets, relative the pre-matched cohort. To sum, this manuscript provides a synopsis of how to notate and use causal methods for categorical treatments.

Difference-based methods do not require estimating the mean function in nonparametric regression and are therefore popular in practice. In this paper, we propose a unified framework for variance estimation that combines the linear regression method with the higher-order difference estimators systematically. The unified framework has greatly enriched the existing literature on variance estimation that includes most existing estimators as special cases. More importantly, the unified framework has also provided a smart way to solve the challenging difference sequence selection problem that remains a long-standing controversial issue in nonparametric regression for several decades. Using both theory and simulations, we recommend to use the ordinary difference sequence in the unified framework, no matter if the sample size is small or if the signal-to-noise ratio is large. Finally, to cater for the demands of the application, we have developed a unified R package, named VarED, that integrates the existing difference-based estimators and the unified estimators in nonparametric regression and have made it freely available in the R statistical program http://cran.r-project.org/web/packages/.

Comparisons of different treatments or production processes are the goals of a significant fraction of applied research. Unsurprisingly, two-sample problems play a main role in statistics through natural questions such as “Is the the new treatment significantly better than the old?” However, this is only partially answered by some of the usual statistical tools for this task. More importantly, often practitioners are not aware of the real meaning behind these statistical procedures. We analyze these troubles from the point of view of the order between distributions, the stochastic order, showing evidence of the limitations of the usual approaches, paying special attention to the classical comparison of means under the normal model. We discuss the unfeasibility of statistically proving stochastic dominance, but show that it is possible, instead, to gather statistical evidence to conclude that slightly relaxed versions of stochastic dominance hold.